Machine learning stock selection: Evidence from the South African factor zoo

Abstract

This study examines whether machine learning can predict Johannesburg Stock Exchange stock returns using South African factor zoo features. Six models are tested in an expanding-window, walk-forward design, with portfolio performance evaluated across alternative weighting schemes and factor-spanning regressions. Ensemble tree-based models, particularly Random Forest, XGBoost and LightGBM, deliver the strongest out-of-sample performance, especially under more intensive training and linear rank weighting. Market-capitalisation weighting weakens alpha. Although machine-learning portfolios generate meaningful alphas under Fama–French models, the inclusion of momentum materially reduces alpha, highlighting momentum’s dominance in South African equity returns.